YOLO Based Breast Masses Detection and Classification in Full-Field Digital Mammograms

Area of Science:

• Medical Imaging
• Bioinformatics
• Deep Learning

Background:

• Mammography is critical for breast cancer screening.
• Human interpretation of mammograms is time-consuming, costly, and error-prone.
• Deep learning, particularly Convolutional Neural Networks (CNNs), shows promise in medical image analysis.

Purpose of the Study:

• To develop an automated computer-aided diagnosis (CAD) system for breast mass detection and classification.
• To evaluate the performance of You Only Look Once (YOLO) architectures for this task.
• To compare YOLO with other feature extractors like ResNet and Inception.

Main Methods:

• An end-to-end CAD system using YOLO for mass detection and classification in mammograms.
• Preprocessing DICOM images without data loss.
• Utilizing YOLO-V3 with k-means clustered anchors for improved detection accuracy.
• Comparing YOLO's classification performance with ResNet and InceptionV3 feature extractors.

Main Results:

• YOLO-V3 achieved 89.4% mass detection rate on INbreast mammograms.
• Classification accuracy for benign and malignant masses was 94.2% and 84.6%, respectively.
• Replacing YOLO's classifier with ResNet and InceptionV3 yielded accuracies of 91.0% and 95.5%.

Conclusions:

• The YOLO-based system significantly impacts breast mass detection and classification.
• K-means clustering for anchor box generation in YOLO-V3 enhances detection of challenging masses.
• Dataset augmentation strategies are crucial, with training set augmentation being most effective for realistic scenarios.